List of Papers Browse by Subject Areas Author List
Abstract
Weight-bearing cone-beam computed tomography (WB-CBCT) enables physiological load-bearing imaging but suffers from motion artifacts caused by involuntary patient tremors. Existing software-based approaches typically address image-domain restoration and motion estimation as separate problems, often leading to over-smoothed reconstructions or suboptimal convergence. To address these problems, we propose Co-MoCo, a collaborative motion compensation framework that integrates image-domain priors with physical consistency through an end-to-end differentiable pipeline. The framework employs a two-stage optimization: Stage I utilizes a Motion-Aware Dual-Decoder Network (MADN), trained via Optimal Transport Conditional Flow Matching (OT-CFM), to generate an artifact-free prior. In Stage II, this prior is integrated into an end-to-end differentiable correction loop to iteratively refine 6-degree-of-freedom (6-DoF) rigid motion trajectories. This synergistic design enables image restoration and motion estimation to mutually regularize each other. Extensive experiments on simulated datasets demonstrate that Co-MoCo significantly outperforms competing methods, achieving SSIM of 0.9849 under mild motion and 0.9152 under severe motion, with performance gains of up to 0.0528. Real-world validation on human and porcine hind limbs further confirms its superior robustness in artifact suppression and detail preservation for high-quality WB-CBCT imaging. Code is available at https://github.com/TomSHINING/CoMoCo.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2489_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TomSHINING/CoMoCo
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LinTon_CoMoCo_MICCAI2026,
author = { Lin, Tong AND Wang, Xi AND Xi, Yan AND Tang, Hao AND Chen, Yang AND Zhang, Yikun},
title = { { Co-MoCo: Collaborative Learning Enabled Motion Compensation for Weight-Bearing CBCT } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.they propose a motion compensation framework for weight-bearing cone-beam computed tomography. This framework integrates image-domain priors with physical consistency through an end-to-end differentiable reconstruction pipeline. 2.they design a Motion-Aware Dual-Decoder Network (MADN) in Stage I of their pipeline. This network is trained using Optimal Transport Conditional Flow Matching to effectively generate an artifact-free prior for images corrupted by motion. 3.they introduce a synergistic optimization approach that overcomes the limitations of isolating image restoration from motion estimation. In Stage II, the artifact-free prior from the MADN is integrated into a differentiable correction loop to iteratively refine 6-degree-of-freedom rigid motion trajectories.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The core innovation of this paper is the introduction of a collaborative motion compensation framework that unites image-domain priors with physical consistency in an end-to-end differentiable pipeline. 2.Comprehensive experiments and evaluation: The authors did not stop at simulated data; they tested the method on a real-world WB-CBCT system with human volunteers.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.While the paper presents a real-world validation on human volunteers and porcine hind limbs , the evaluation is purely qualitative. 2.The proposed Co-MoCo framework relies on an iterative optimization pipeline (Stage II) that updates motion parameters over 100 iterations using a differentiable cone-beam backprojector and a neural network. However, the authors do not report the memory footprint or the time required to reconstruct a single 3D volume on their NVIDIA RTX 4090 setup 3.In Section 2.2, the authors describe the Motion-Aware Dual-Decoder Network (MADN) and state that it takes a motion intensity map ($\sigma_{jit}$) as input. They note that this map is “derived from a pre-trained discriminator that quantifies the severity of motion”. The paper omits the architectural details, training methodology, and data sources used for this “pre-trained discriminator. “
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Major factors supporting the acceptance: Clinical Relevance & Motivation: The paper tackles a highly relevant and challenging problem in musculoskeletal imaging—motion artifacts in Weight-bearing CBCT caused by involuntary tremors. Solving this has direct clinical utility.
Methodological Novelty
Major factors limiting a higher score: Missing Computational/Runtime Analysis: The iterative nature of the differentiable backprojection (Stage II requires 100 iterations) raises significant concerns regarding computational efficiency and memory overhead. The authors completely omit runtime metrics. For a method targeting clinical application, assessing the time required to reconstruct a single 3D volume is critical.
Lack of Quantitative Real-World Metrics: While the visual results on human and porcine data (Figure 4) are encouraging, the evaluation of the real-world porcine data—which explicitly had a stationary reference—is entirely qualitative. The absence of quantitative metrics (like SSIM, PSNR, or MSE) on this real data makes it difficult to objectively verify the method’s effectiveness outside of simulated environments.Incomplete
Technical Details: Crucial details regarding the architecture and training of the “pre-trained discriminator” (used to extract the motion intensity map $\sigma_{jit}$) are omitted, which hinders reproducibility. Furthermore, the reliance on a purely rigid 6-DoF assumption ignores the reality of non-rigid soft tissue deformations during patient tremors.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
All my concerns have been resolved.
Review #2
- Please describe the contribution of the paper
This paper presents Co-MoCo, a collaborative motion compensation framework for weight-bearing CBCT. The method targets motion artifacts in WB-CBCT caused by unstable standing posture, tremor, or fatigue, and attempts to address the limitation that image-domain restoration and motion parameter estimation are often treated separately. Specifically, the proposed framework first employs a Motion-Aware Dual-Decoder Network to learn an artifact-free image prior, and then embeds this prior into an end-to-end differentiable motion correction loop to iteratively optimize a 6-DoF rigid motion trajectory. Experimental results show that the method outperforms multiple image restoration and motion compensation baselines under both mild-motion and severe-motion settings, while also demonstrating promising artifact suppression and detail recovery on real human volunteer and porcine hind-limb data. Overall, the manuscript is clearly written, the methodological logic is complete, and the figures are intuitive and easy to follow.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
(1) The topic is highly practical and addresses a current clinical pain point. The paper focuses on motion artifacts in weight-bearing CBCT, which is a clinically relevant problem in orthopedic and musculoskeletal rehabilitation scenarios. The introduction clearly explains the clinical value of WB-CBCT for weight-bearing imaging of the knee, foot/ankle, and lower extremities, as well as its susceptibility to involuntary patient motion due to the relatively long scanning time. Therefore, the work is well motivated and has a clear application background. (2) The method is well designed and demonstrates a meaningful degree of innovation, with a clear collaborative relationship between its components. The paper attempts to let image restoration and motion parameter estimation mutually reinforce each other within a unified framework, which is more valuable than using restoration-only or motion-estimation-only strategies. In particular, under severe motion, single-module methods are clearly limited, whereas the joint framework is still able to maintain high image quality, suggesting that the proposed design brings real synergistic benefits. (3) The quantitative experiments are comprehensive. The study includes both mild-motion and severe-motion settings and compares the proposed method against multiple image restoration and motion compensation approaches, making the results readable and convincing. (4) The real-data examples enhance the credibility of the method. In addition to simulation experiments, the authors also validate the framework on a real WB-CBCT system using human volunteer data and porcine hind-limb data. Although this part is still mainly qualitative, it nevertheless shows that the method is not limited to a purely simulated environment and has a certain degree of practical feasibility. (5) The proposed method not only achieves better numerical results than the competing methods, but also shows more intuitive improvements in residual blur reduction, streak artifact suppression, and preservation of high-frequency details. For motion artifact correction tasks, metric improvements alone are sometimes insufficient to demonstrate practical value, and the visual improvements shown in this paper make the conclusions more convincing.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(1) The real-data validation remains largely qualitative and lacks further quantitative analysis. Although the human volunteer and porcine hind-limb results strengthen the practical relevance of the paper, this part currently relies mainly on visual inspection and lacks a more systematic quantitative evaluation. As such, the generalization ability and stability of the method in real clinical scenarios have not yet been fully established. (2) The assumption on the motion model is somewhat idealized. The paper mainly models patient motion as a 6-DoF rigid trajectory. While this assumption may be reasonable for some motion patterns in lower-limb WB-CBCT, in real scenarios, especially when soft tissue or local limb posture changes are more pronounced, more complex local non-rigid deformations may still occur. The paper does not yet discuss the applicability boundary of this assumption in sufficient depth, and therefore the robustness of the method under more complex motion patterns remains to be demonstrated. (3) The discussion of runtime efficiency and deployment cost is insufficient. The manuscript does not explicitly report the inference time or optimization time per case. Given that the proposed framework may require iterative optimization of motion parameters within a differentiable reconstruction loop, runtime and computational cost will be important considerations for future clinical deployment. The current version is somewhat limited in this respect. (4) Some baselines perform relatively poorly under severe motion, but the paper does not further analyze the reasons for this. Several image-domain baselines show substantial performance degradation in the severe-motion setting, yet the manuscript does not discuss why these baseline results are poor. This weakens the completeness of the comparative study to some extent.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, I am inclined toward weak accept, because I believe this paper addresses a clinically meaningful and practically valuable problem, and the proposed framework is also technically well motivated. In addition, the manuscript is generally clear and easy to follow, and the qualitative and quantitative results together make the work convincing as a promising solution for motion artifact suppression in WB-CBCT.However, I do not give a full accept for two main reasons. First, the validation on real data still relies primarily on qualitative analysis, and the evidence for robustness and generalization in realistic clinical settings remains limited. Second, the method appears difficult to reproduce in practice, since the framework is relatively complex and the paper does not indicate that the implementation code will be publicly available. As a result, reproducing the full pipeline may require substantial effort.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I appreciate the authors’ efforts in providing additional clarification and supplementary quantitative results in the rebuttal. However, I still feel that several issues have not been fully addressed. The quantitative validation on real data appears to remain limited and may not be sufficient to establish the generalization and stability of the method in practical clinical scenarios. The discussion of the 6-DOF rigid body motion assumption is reasonable, but it also shows that it is not applicable to more complex local or non-rigid body motion. In addition, although the authors now report computational cost, the reported runtime suggests that clinical deployment may still be challenging.
Review #3
- Please describe the contribution of the paper
The paper proposes Co-MoCo, a motion compensation framework for Weight-bearing CBCT (WB-CBCT). It addresses motion artifacts through a two-stage process: first, training a Motion-Aware Dual-Decoder Network (MADN) using Optimal Transport Conditional Flow Matching (OT-CFM) to serve as an artifact-free prior; second, utilizing this frozen prior within a differentiable correction loop to iteratively optimize the 6-DoF rigid motion parameters (P-Matrix) based on physical consistency. Extensive evaluations demonstrate that the proposed method achieves state-of-the-art performance across both simulated motion datasets and real-world clinical scans.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper innovatively couples an image-domain generative prior with a physical forward model, solving the motion-corrupted inverse problem by iteratively optimizing 6-DoF rigid motion parameters to overcome the over-smoothing limitations of pure image-domain methods. 2.The paper introduces Optimal Transport Conditional Flow Matching (OT-CFM) to train a dual-decoder network (MADN), which serves as a high-quality, artifact-free manifold regularization constraint for the subsequent joint optimization of physical parameters. 3.The paper demonstrates the clinical feasibility and effectiveness of the proposed framework through comprehensive validation on both simulated motion datasets and real-world clinical scans.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The motivation for selecting Optimal Transport Conditional Flow Matching (OT-CFM) as the artifact-free prior, rather than other generative models, is insufficiently justified. Additionally, the authors need to clarify whether the identical flow matching setup is applied across different motion parameters, and rigorously explain how a completely frozen MADN can consistently guarantee reliable prior guidance during the iterative optimization process. 2.The strategy of alternating optimization between image domains and motion parameters has been extensively explored in previous literature. The authors should properly cite these prior works and explicitly discuss the methodological differences and specific advantages of the proposed framework over existing alternating optimization schemes. 3.To convincingly demonstrate its effectiveness, the experimental section must include a quantitative comparison with at least one state-of-the-art baseline method that jointly optimizes image and motion parameters. 4.The authors are strongly encouraged to open-source their codebase and release the real-world clinical dataset to facilitate reproducibility and benefit related research within the medical imaging community. 5.The manuscript lacks structural cohesion in the methodology section. The logical flow and transitions between Sections 2.1, 2.2, and 2.3 are somewhat disjointed and require further elaboration to present a unified narrative. 6.The paper omits critical computational efficiency metrics. The authors should report and compare the network parameter counts and running times against the baseline methods to verify the real-time clinical feasibility of the proposed pipeline.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the proposed physics-guided framework demonstrates strong clinical feasibility and innovative design, the manuscript still requires crucial clarifications regarding its theoretical motivation, joint-optimization baseline comparisons, and computational efficiency metrics.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The author has addressed most of my concerns.
Author Feedback
We thank AC for the insightful summary. We thank R2 and R3 for accepting our paper as “tackles a highly relevant and challenging problem”, “addresses a clinically meaningful and practically valuable problem”, and thank R4 for appreciating our work as “strong clinical feasibility and innovative design”. Below, we provide point-by-point responses.
AC-Q1&R2-Q1&R3-Q1: Lack of quantitative analysis on real data Acquiring GT in WB-CBCT scans is hard due to inevitable patient motion, though motion-tracking devices [1] could provide reliable references in future work. To provide a quantitative analysis on real data, we report results on the porcine case in Fig. 4.Our approach achieves lower MSE (6691.25 vs. 7412.84/8688.25/10147.37), higher SSIM (0.8411 vs. 0.8132/0.7881/0.7171), and higher PSNR (35.49 vs. 34.55/33.08/31.44) than Thies’s, Uformer, and FDK. [1] Maier J, et al. Inertial measurements for motion compensation in weight-bearing cone-beam CT of the knee.
AC-Q2&R2-Q2&R3-Q3&R4-Q6: Lack of computational cost information Our method, run for 100 iterations with 40 control points on an NVIDIA RTX 4090 (projection size 400×1024×1536; volume 80×512×512/80×256×256), requires 9055/2256 sec and 14.0/12.9 GB GPU memory, while Thies’s [25] takes 8840/2203 sec and 15.3/13.7 GB GPU memory.
AC-Q3&R2-Q4&R3-Q2: Validity of the 6-DOF assumption The 6-DOF rigid motion assumption is widely adopted for bone-dominant motion. However, lower-limb motion can be influenced by joint articulation and soft-tissue deformation. On real data, performing separate optimizations for thigh, knee, and leg can mitigate this issue, as shown in Fig. 4.Above limitations and potential of non-rigid modeling will be discussed in revision.
AC-Q4&R4-Q1: Intuition behind artifact-free prior 1)OT-CFM enables stable and continuous trajectory learning by leveraging few-step transport estimation, providing faster inference than diffusion-based models. As shown in Tables 1 and 2, it also outperforms competing NN methods. 2) The MADN maintains a unified flow framework and uses varying degrees of motion information through a pre-trained discriminator, thereby performs effectively across different motion intensities. Additional ablation studies were omitted due to constraints.
AC-Q5&R3-Q4: Concerns about baseline comparisons Under severe motion, image quality is highly degraded. Purely image-domain NN provides multiple plausible solutions for a single voxel, leading to excessive blurring. The proposed framework utilizes real projections and scan geometry as strong physical constraints through a differentiable P-Matrix back-projector, thus iteratively narrowing solution space toward a physically consistent state.
R2-Q4&R3-Q4&R4-Q4: Concerns about reproducibility We promise to release code once acceptance. Clinical datasets cannot be shared due to hospital privacy.
R2-Q3: Lack of implementation details of pre-trained discriminator Motion maps are generated by a lightweight U-Net trained on simulated motion dataset (vary degrees of motion) to provide motion intensity information for the MADN. More details will be found in our released code.
R4-Q2&Q3: Discussion and comparison with existing alternating optimization schemes Previous alternating schemes primarily rely on costly, view-constrained forward-projector, suboptimal regularizers such as TV and VIF and on scarce high-quality CAD (on head data). In contrast, we utilize a differentiable P-Matrix back-projector to enable strong physical constraints and high-quality OT-CFM guidance. As noted in Table 1, we have included a comparison with the state-of-the-art alternating optimization scheme by Thies [25]. More comprehensive comparisons will be conducted in future work.
R4-Q5: Lack of structural cohesion in methodology Our methodology is organized as: Problem Formulation (Sec. 2.1), Prior Generation (Sec. 2.2), Physical Modeling (Sec. 2.3), Overall Architecture (Sec. 2.4). We will improve structural coherence in revision.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The authors proposed Co-MoCo, a collaborative motion compensation framework that integrates image-domain priors with physical consistency through an end-to-end differentiable pipeline. Stage 1 contains a Motion-Aware Dual-Decoder Network (MADN) trained via Optimal Transport Conditional Flow Matching (OT-CFM) to generate an artifact-free prior. In Stage 2 this prior is integrated into a correction loop to iteratively refine 6-DoF rigid motion trajectories.
Reviewers all gave weak acceptance/borderline scores. The motivation behind image-domain priors with physical consistency is widely acknowledged, and the experiments are comprehensive with real-patient data. However, it’s noted that common concerns include (1) qualitative evaluation on real-patient data, which might be reasonable since there is no motion GT, but the authors might need to further discuss this and include potential future directions with motion-tracking devices to get GT; (2) lacking computational cost information since there are iterative optimization loops; (3) the validity of the 6-DOF motion trajectory assumption; (4) the intuition behind artifact-free prior selection and (5) concerns regarding baseline comparisons. The authors are encouraged to address these concerns and provide other missing details as required by the reviewers.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Thank the authors for providing this detailed response. Most of the concerns were sufficiently addressed, including real data quantitative analysis and runtime report. The authors are encouraged to discuss the potential of non-rigid motion modeling in the camera-ready version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After rebuttal, the paper leans toward acceptance. The reviewers agreed that the problem is clinically important and that the proposed Co-MoCo framework offers an interesting integration of learned image priors with physics-based motion correction for WB-CBCT. The rebuttal addressed several major concerns by providing runtime and memory estimates, clarification of the 6-DOF assumption, additional details on the motion-intensity discriminator, and discussion of prior alternating-optimization methods. Two reviewers were satisfied by the rebuttal and now recommend accept. Reviewer 2 maintains a reject recommendation, citing limited real-data validation, long runtime, and the remaining limitations of the rigid-motion model. These are valid concerns and should be acknowledged, particularly the high computational cost and limited quantitative real-data evidence. However, given the novelty and clinical relevance of the framework, the improved clarification after rebuttal, and the overall positive reviewer balance, I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have responded to most of the reviewers’ comments. While one reviewer still has concerns about the evaluation and feasibility of clinical deployment, I think the authors should expand their discussion on the limitations in the final version.
